arXiv:2605.06684v1 Announce Type: new Abstract: Tree-involved crashes represent a critical subset of run-off-road (ROR) collisions, often resulting in fatal or severe injuries due to high-energy impacts. This study develops a comprehensive analytical framework to identify and quantify risk factors contributing to crash severity in tree-involved collisions using the Crash Report Sampling System (CRSS) database spanning 2020-2023. The modeling framework follows a multi-step process. First, a machine learning based classification model (CatBoost) identifies key factors associated with binary crash injury severity (KA: fatal or incapacitating injury versus BC: non-incapacitating or possible injury). Second, SHapley Additive exPlanations (SHAP) tool is used to quantify and visualize the marginal effects of top influential factors on crash severity. Third, a binary logistic regression model estimates factor effects and validates SHAP-derived importance measures. Finally, SHAP interaction plots examine the combined effects of key contributing factors. Results reveal restraint non-use as the most influential predictor, with unrestrained occupants nearly three times more likely to experience severe outcomes due to ejection risk. Vehicle age, speeding violations, and driver impairment demonstrate substantial effects, reflecting reduced crashworthiness, increased impact forces, and reduced control capabilities. Critical interactions emerge between lighting conditions and vehicle age, speeding and lighting conditions, restraint use and vehicle age, and road surface and speeding, demonstrating additive risk effects with specific interactions. These findings provide critical insights for targeted safe system-based interventions, including enhanced seat belt enforcement, speed management in reduced visibility conditions, and vehicle fleet modernization.

arXiv:2605.06758v1 Announce Type: new Abstract: Relative spatial relations provide a compact representation of spatial structure and are fundamental to relative spatial reasoning in 3D layout generation. Recent works leverage Multimodal Large Language Models (MLLMs) to infer such relations, but the inferred relations are often unreliable and are typically handled with post-hoc heuristics. In this paper, we propose R$^3$L, a general framework that improves the reliability and consistency of relative spatial reasoning for 3D layout generation. Our key motivation is that multi-hop reasoning requires repeated reference-frame transformations, which accumulate errors in inferred relations and lead to semantic and metric drift. To mitigate this, we propose invariant spatial decomposition to break coupled relation chains, and consistent spatial imagination to promote self-consistency through an imagine-and-revise loop. We further introduce supportive spatial optimization to ease pose optimization via global-to-local coordinate re-parameterization. Extensive experiments across diverse scene types and instructions demonstrate that R$^3$L produces more physically feasible and semantically consistent layouts. Notably, our analysis shows that resolving frame-induced inconsistencies is crucial for reliable multi-hop relative spatial reasoning. The code is available at https://github.com/Neal2020GitHub/R3L.

arXiv:2605.06683v1 Announce Type: new Abstract: Transformer-based large language models are in some respects limited by the quadratic time and space computational complexity of attention. We introduce the Toeplitz MLP Mixer (TMM), a transformer-like architecture that swaps attention for triangular-masked Toeplitz matrix multiplication over the sequence dimension resulting in $\mathcal{O} (dn \log n)$ time and $\mathcal O(dn)$ space complexity during training and $\mathcal O(dn)$ time and space at inference prefill. Despite the lack of sophisticated input modulation or state maintenance present in other sub-quadratic architectures, TMMs yield greater training efficiency in terms of loss achieved per compute and device memory. We demonstrate that TMMs are capable of retaining more input information resulting in improved copying ability, which we argue results from a lack of architectural biases. Consistent with higher input information retention, TMMs exhibit superior information retrieval and in-context learning benchmark accuracy compared to comparable architectures. We conclude with an analysis from the perspective of operator index theory and show that, counterintuitively, trained Toeplitz layers of causal non-invertible models are more likely to be invertible or nearly so than models that are actually invertible over their inputs.

arXiv:2605.06681v1 Announce Type: new Abstract: A hierarchical ensemble pipeline is introduced to address anomaly detection in multivariate telemetry data provided by European Space Agency (ESA). The method integrates shapelet-based and statistical feature extraction, per-channel modeling, intra-channel stacking, and a final cross-channel aggregation. The pipeline is trained and validated using time-series cross-validation and two-level masking strategies to prevent information leakage. Results on the European Space Agency Anomaly Detection Benchmark (ESA-ADB) challenge demonstrate strong generalization, highlighting the effectiveness of hierarchical modeling in detecting subtle anomalies in realistic satellite telemetry.

arXiv:2605.06673v1 Announce Type: new Abstract: Aggregate metacognitive quality scores mask within-model variation across MMLU benchmark domains. We administered 1,500 MMLU items (250 per domain, under an a priori six-domain grouping) to 33 frontier LLMs from eight model families and computed Type-2 AUROC per model-domain cell using verbalized confidence (0-100). Total observations: 47,151. Every model with above-chance aggregate monitoring showed non-trivial domain-level variation. Applied/Professional knowledge was reliably the easiest benchmark domain to monitor (mean AUROC = .742, ranked top-2 in 21 of 33 models); Formal Reasoning and Natural Science were reliably the hardest (one of the two ranked bottom-2 in 27 of 33 models). The three middle domains were statistically indistinguishable (Kendall's W = .164). A subject-level coherence analysis (within-domain similarity ratio = 0.95) confirms the six-domain grouping is a pragmatic benchmark taxonomy, not a validated latent construct. Within-family profile-shape clustering is significant for Anthropic, Google-Gemini, and Qwen (permutation p < .0001) but not DeepSeek, Google-Gemma, or OpenAI. Gemma 4 31B showed a +.202 AUROC improvement over Gemma 3 27B. Three models classified Invalid on binary KEEP/WITHDRAW probes produced normal profiles under verbalized confidence, confirming probe-format specificity. Bootstrap 95% CIs on 198 cells have median width .199. Split-half aggregate stability r = .893; profile-level split-half is weaker (grand median r = .184). These results show stable benchmark-domain variation obscured by aggregate metrics, and support benchmark-stage domain screening as a step before deployment in specific application areas.

arXiv:2605.06832v1 Announce Type: new Abstract: Accurately understanding the intent behind speech, conversation, and writing is crucial to the development of helpful Large Language Model (LLM) assistants. This paper introduces IntentGrasp, a comprehensive benchmark for evaluating the intent understanding capability of LLMs. Derived from 49 high-quality, open-licensed corpora spanning 12 diverse domains, IntentGrasp is constructed through source datasets curation, intent label contextualization, and task format unification. IntentGrasp contains a large-scale training set of 262,759 instances and two evaluation sets: an All Set of 12,909 test cases and a more balanced and challenging Gem Set of 470 cases. Extensive evaluations on 20 LLMs across 7 families (including frontier models such as GPT-5.4, Gemini-3.1-Pro, and Claude-Opus-4.7) demonstrate unsatisfactory performance, with scores below 60% on All Set and below 25% on Gem set. Notably, 17 out of 20 tested models perform worse than a random-guess baseline (15.2%) on Gem Set, while the estimated human performance is ~81.1%, showing substantial room for improvement. To enhance such ability, this paper proposes Intentional Fine-Tuning (IFT), which fine-tunes the models on the training set in IntentGrasp, yielding significant gains of 30+ F1 points on All Set and 20+ points on Gem Set. Tellingly, the leave-one-domain-out (Lodo) experiments further demonstrate the strong cross-domain generalizability of IFT, verifying that it is a promising approach to substantially enhancing the intent understanding of LLMs. Overall, by benchmarking and boosting intent understanding ability, this study sheds light on a promising path towards more intentional, capable, and safe AI assistants for human benefits and social good.

arXiv:2605.06903v1 Announce Type: new Abstract: Large language models are now embedded in everyday writing workflows, making reliable AI-generated text detection important for academic integrity, content moderation, and provenance tracking. In practice, however, a detector must do more than achieve high aggregate AUROC on clean, in-distribution human and AI text: it should remain robust to attacks and adversarial rewrites, transfer to unseen generators and domains, and operate at low false-positive rates (FPR). Most existing detectors optimize a single AI/Human objective, giving the representation little incentive to learn generator, attack, or domain structure once the binary task saturates. We introduce MELD (Multi-Task Equilibrated Learning Detector), a deployable detector for AI-generated text that enriches binary detection with auxiliary supervision. MELD attaches generator-family, attack-type, and source-domain heads to a shared encoder, and balances the four losses with learned homoscedastic uncertainty weights. To improve robustness, an EMA teacher predicts on clean inputs while an attack-augmented student is distilled toward the teacher. MELD further uses a hard-negative pairwise ranking loss to enlarge the score margin between AI-generated texts and the most confusable human texts. At inference, all auxiliary heads are discarded, giving MELD the same interface and cost as a standard detector. On the public RAID leaderboard, MELD is the strongest open-source detector and is competitive with leading commercial models, especially under attack and at low FPR. Across standard held-out benchmarks, MELD matches or outperforms supervised baselines. We further introduce MELD-eval, a held-out evaluation pool built from recent chat models released by four major LLM providers. Without additional finetuning, MELD achieves 99.9% TPR at 1% FPR on MELD-eval, while many baselines degrade sharply.

arXiv:2605.06676v1 Announce Type: new Abstract: Long-context inference in Large Language Models (LLMs) is bottlenecked by the linear growth of Key-Value (KV) cache memory. Existing KV cache compression paradigms are fundamentally limited by heuristics: heuristic budgeting relies on statistical priors rather than task objectives, causing resource misallocation, while heuristic selection relies on coupled query-key interactions or static inductive biases (e.g., attention sinks). To address this limitation, we introduce LKV (Learned KV Eviction), which formulates KV compression as an end-to-end differentiable optimization problem. LKV integrates LKV-H to learn task-optimized global budgets, and LKV-T to derive intrinsic KV importance without materializing attention matrices. This design bypasses heuristic proxies, strictly aligning compression with task objectives. Extensive evaluations demonstrate that LKV achieves state-of-the-art performance on both LongBench and RULER benchmarks at high compression rates. In particular, on LongBench, LKV achieves near-lossless performance with only 15\% KV cache retention. Crucially, our analysis identifies learned budgeting as the dominant driver of fidelity, demonstrating that data-driven allocation is essential to overcome the limitations of hand-crafted heuristics.

arXiv:2605.06675v1 Announce Type: new Abstract: Large language models cache all previously computed key-value (KV) pairs during generation, and this KV cache grows linearly with sequence length, making it a primary memory bottleneck for serving. Quantizing the KV cache to fewer bits reduces this cost, yet all current quantizers assign the same bit-width to every attention head, ignoring the large variation in head importance. A natural idea is to allocate more bits to important heads and fewer to the rest. We show, however, that such mixed-precision allocation has a hidden pitfall: each quantizer follows a different distortion curve D(b)=alpha*beta^{-b}, and the decay rate beta varies from 3.6 to 5.3 across quantizer designs. Applying one quantizer's distortion model to another inverts the allocation order and makes performance worse than uniform quantization. We call this failure mode distortion model mismatch and propose RateQuant to resolve it. RateQuant fits a per-quantizer distortion model from a small calibration set, then solves the resulting bit-allocation problem in closed form via reverse waterfilling from rate-distortion theory. On Qwen3-8B at 2.5 average bits, calibrated RateQuant reduces KIVI's perplexity from 49.3 to 14.9 (70% reduction) and improves QuaRot by 6.6 PPL. The entire calibration takes 1.6 s on a single GPU and adds zero overhead at inference time.

arXiv:2605.06919v1 Announce Type: new Abstract: Large language models have demonstrated impressive retrieval-augmented capabilities. However, a crucial area remains underexplored: their ability to appropriately adapt responses to the certainty of the retrieved information. It is a limitation with real consequences in high-stakes domains like medicine and finance. We evaluate eight LLMs on their context-certainty obedience, measuring how well they adjust responses to match expressed context certainty. Our analysis reveals systematic limitations: LLMs struggle to recall prior knowledge after observing an uncertain context, misinterpret expressed certainties, and overtrust complex contexts. To address these, we propose an interaction strategy combining prior reminders, certainty recalibration, and context simplification. This approach reduces obedience errors by 25% on average, without modifying model weights, demonstrating the efficacy of interaction design in enhancing LLM reliability. Our contributions include a principled evaluation metric, empirical insights into LLMs' uncertainty handling, and a portable strategy to improve context-certainty obedience across diverse LLMs.

arXiv:2605.06901v1 Announce Type: new Abstract: Large Language Models (LLMs) are increasingly shaping the private and professional lives of users, with numerous applications in business, education, finance, healthcare, law, and science. With this rise in global influence comes greater urgency to build, evaluate, and deploy these systems in a manner that prioritizes not only technical capabilities but also human priorities. This work presents a framework for developing Human-Centered Large Language Models (HCLLMs), which integrates perspectives from Natural Language Processing (NLP), Human-Computer Interaction (HCI), and responsible AI. Considering the ethics, economics, and technical objectives of language modeling, we argue that model developers need to address human concerns, preferences, values, and goals, not only during a cursory post-training stage, but rather with rigor and care at every stage of the pipeline. This paper offers human-centered insights and recommendations for developers at each stage, from system design to data sourcing, model training, evaluation, and responsible deployment. Then we conclude with a case study, applying these insights to understand the future of work with HCLLMs.

arXiv:2605.06897v1 Announce Type: new Abstract: The rise of Internet of Things (IoT) devices in the physical world necessitates voice-based interfaces capable of handling complex user experiences. While modern Large Language Models (LLMs) already demonstrate strong tool-usage capabilities, modeling real-world IoT devices presents a difficult, understudied challenge which combines modeling spatiotemporal constraints with speech inputs, dynamic state tracking, and mixed-initiative interaction patterns. We introduce MIST (the Multimodal Interactive Speech-based Tool-calling Dataset), a synthetic multi-turn, voice-driven code generation task that operates over IoT devices. We find that there is a significant gap between open- and closed-weight multimodal LLMs on MIST, and that even frontier closed-weight LLMs have substantial headroom. We release MIST and an extensible data generation framework to build related datasets in order to facilitate research on mixed-initiative voice assistants which reason about physical world constraints.

Fortune: Q&A with Qualcomm CEO Cristiano Amon on 2026 as “year of agents”, the end of the smartphone-centric world, 6G turning humans into “walking cameras”, and more  —  Cristiano Amon's pitch is that the smartphone-centric world Qualcomm helped build is coming to an end.

Victoria Aleksandrova/Unsplash, CC BY Unemployment in Australia is currently at 4.3% – below average for the past century. But last week, the Reserve Bank forecast unemployment will slowly climb from next year on, rising to 4.7% by mid-2028 – and possibly higher if the Middle East war drags on. That forecast came after a long-running survey of 3,600 Australians – taken in the early weeks of the US-Israel war on Iran – found people thought their chances of losing their job had risen to 26.8%. Job insecurity fears haven’t been this high since 2020’s COVID-19 lockdowns, when unemployment hit 6.4%. If worrying about keeping your job has been keeping you up at night, you’re far from alone. But there are evidence-based things we can do – at an individual, organisational and government level – to manage job insecurity better in uncertain times. How job stress hits our health and even our personality Job insecurity can take a heavy toll on your mental and physical health. There’s strong evidence from meta analyses – research where all the studies in that area are pulled together – that people are significantly more dissatisfied with work when experiencing job insecurity. The same study found job insecurity can also affect workers’ commitment to their organisation and undermine their job performance. There’s also some evidence job insecurity can increase workplace bullying. When people feel really insecure, they’re more likely to lash out at others. The same study showed workers who’ve been bullied tend to feel more insecure in their work. Our 2020 study, tracking 1,046 Australians over nearly a decade, suggested a prolonged period of job insecurity could even change people’s personalities: making them less emotionally stable, less agreeable and less conscientious. 5 ways to turn fear into action It’s easy to fall into the trap of unhealthy coping strategies when you’re stressed about work. These can include drinking too much alcohol or “ruminating”: allowing worries to go around and around in your brain. While these can feel hard to resist, they usually leave you feeling worse. There are healthier coping strategies, backed by research evidence. Actively planning your career: This sounds obvious, but it can make a big difference. In a study my colleagues and I did of more than 200 Europeans on short-term contracts, we found that insecurity increased the closer people got to the end of their contracts – unless people took proactive steps. Workers who took the initiative to engage in career planning and talk to supervisors and others about how to boost their skills did not experience job insecurity, even as their short-term contracts got close to ending. Seeking social support: reaching out to other people – family, friends, colleagues or your boss – is a well-established way to help with work stressors, such as job insecurity. For instance, a meta analysis of three decades of data from 39 countries found social support can reduce feelings of job insecurity. Building a network with informal mentors: You don’t need a formal mentor program to do this. In fact, past evidence has shown informal mentoring (usually someone you’ve approached yourself) can be more powerful than formal mentoring. And seek out more than one mentor: one person might help on digital skills, another might have advice on career planning. Working on in-demand skills: If you know you have gaps in your skills, such as if you’re worker with no experience with artificial intelligence (AI), start taking small steps to fill those gaps. There are a wide range of free online courses you can do from places such as EdX and Coursera, through to relatively affordable “micro credential” courses from universities around the world. Building a life beyond work: Research shows job insecurity is more harmful to people for whom work is highly central to their identity. A 2025 Australian study confirmed what decades of research has shown: engaging in other social roles beyond work – like being a parent, friend, or volunteer – can boost your self-esteem and wellbeing. Employers and government should do more too It’s unfair to expect individuals to handle job insecurity on their own. Employers and governments have important roles to play too. Here are just two examples. Better workplace training: Australians are worried about losing work to AI. Federal government research last year found more Australians are secretly using artificial intelligence, often in the absence of clear workplace rules. If organisations want to retain good staff, it’s in their interests to invest in employee training, including in AI – and have clearer policies on its use. Employment data shows the early signs of AI job disruption are already here Government investment in training and welfare: A 2022 study comparing 17 European nations found workers experienced less job insecurity in nations such as Denmark and Sweden, because they invest more in helping people get work, such as plentiful opportunities for re-skilling, training, and career counselling. Generous employment benefits also improve people’s wellbeing through fostering financial security. In Australia, JobSeeker payments to 900,000 unemployed, working-age people still fall well below the poverty line, despite repeated recommendations to lift them higher. Increasing JobSeeker is long overdue. Here’s how we could do it, without breaking the budget There are few jobs for life anymore. But amid so much change, there’s more we can do more to help ourselves, our colleagues, employees and those who are already unemployed feel less job insecurity than Australians currently do. Sharon Kaye Parker receives funding from the Australian Research Council.

Sarah Kessler / New York Times: Corporate lawyers say AI-transcribed meeting notes may not have attorney-client privilege, making otherwise protected conversations discoverable in lawsuits  —  A trendy productivity hack, A.I. note takers are capturing every joke and offhand comment in many meetings.

In this article, we fine-tune the Qwen3.5-0.8B model on the VQA-RAD dataset, which is a question-answering dataset based on radiology images. After training, we carry out inference using the fine-tuned model. The post Fine-Tuning Qwen3.5 appeared first on DebuggerCafe.

Amrisha Agarwal moved to Rishikesh with her daughter, Lavenia, to escape the rat race of city life. Amrisha Agarwal Amrisha Agarwal moved to Rishikesh, a town in the Himalayas, with her daughter in 2019. She opened a café, where she serves dishes to a diverse clientele. Agarwal said the peace and quality of life she found have been instrumental in her daughter's upbringing. This as-told-to essay is based on a conversation with Amrisha Agarwal, a 36-year-old café owner in Rishikesh, India. It has been edited for length and clarity. I was born and raised in Kolkata, a populous city in India's far eastern state of West Bengal. I was married there, gave birth to my daughter there, and built a brand consultancy with my ex-husband. But I've always dreamed of moving out of the city. When I visited Rishikesh, a quiet town in northern India, it was clear that this was the place I'd been searching for. A peaceful mountain escape Agarwal moved to Rishikesh in 2019 to escape a difficult marriage and the city's rat race. Rebecca Conway/Getty Images In 2019, my marriage was not in a good place. That same year, my father died. After his cremation, I came to Rishikesh with his ashes to immerse them in the Ganges River. My first impression of Rishikesh was that it was peaceful and pristine. It had really clean, fresh air, which I wouldn't get in Kolkata or any other city in India. My health, my gut, my digestion — everything improved with better food and cleaner water. As a woman, I felt a lot of freedom here. I didn't have to think about what time it was, I didn't feel unsafe, and I didn't have to be conscious of what I was wearing. That kind of ease is something I had never really experienced before. I couldn't let go of my father immediately. It took me almost 15 days before I could finally say goodbye. Something shifted during that time. I had a return ticket, everything was planned, but I couldn't leave. I remember the taxi was waiting to take me to the airport, and I said, "I'm not going back." Building a café with no expertise After I decided to stay in Rishikesh and moved my daughter and me here permanently, I had no game plan. I just needed a break and space in my relationship. I had not planned to work in Rishikesh, since I already had a company back in Kolkata. For the first few months, I lived off my savings. But in our rented apartment, I started cooking more because my daughter and I were missing the flavors we couldn't find in Rishikesh. We made friends in community spaces, and people started coming over to eat my food. They asked me, "Why not make a space and sell your food?" That's how Bistro de Lavenia started, named after my daughter. When we started in December 2020, I had no expertise in running a food business. I don't know what kind of equipment was needed. We started with just four dishes: two kinds of pizza and two kinds of pasta. Amrisha's café now sells a variety of dishes from around the world. Amrisha Agarwal Five years on, the menu has expanded vastly, with Asian and Indian dishes and some international varieties. We have eight staff members, and I've brought on a friend to help design the menu and handle the food operations. I see Rishikesh's effect on my daughter Amrisha's daughter, Lavenia, has grown up in Rishikesh. Amrisha Agarwal Rishikesh has been instrumental in my daughter's upbringing. When she moved here, she was about 2.5 years old. She's nine now, so practically her entire development and early childhood have happened in Rishikesh. When she got bored, instead of giving her a screen, I would take her out to the Ganges, to the sand, to waterfalls, for hikes, cycling, skating, and archery. Her life has been built around doing things physically, being outdoors, rather than sitting in one place watching television or being on the phone. Rishikesh gets visitors from everywhere because it's India's yoga capital, which is why she's grown up meeting people from all over the world. There was this one time she met a family from Russia in the café, and they were talking about carrots. She only knew carrots to be orange. Then their daughter told her that back home, they had yellow and purple carrots too. It's these kinds of things that make me appreciate that my daughter's education isn't just from books; it's from real interactions, and makes her realize that the world is much bigger than what we are taught. Rishikesh is changing Rishikesh, India's yoga capital, is getting more crowded. Rebecca Conway/Getty Images Rishikesh has grown and changed from the peaceful haven it once was for me. When I started my café, there were maybe five cafés in total on the main street, including mine. You would see a few lights from those buildings, and apart from that, there was pitch darkness at night. Now it's completely different. It feels like Rishikesh has grown a hundred times. The biggest shift happened after the pandemic. A lot of people moved here as remote work became popular, and the number of businesses here suddenly exploded. People saw an opportunity, and everyone started opening cafés. It's started to feel like the city I left behind. I look out of my window instead of seeing the mountains, I see another building, and that's not what I signed up for. I'm planning my next move. I don't know which part of India or the world, but I'd like to be someplace where the next building is at least 10 kilometers away, where we can really pursue slow living once again. Read the original article on Business Insider

<blockquote cite="https://www.nytimes.com/2026/04/14/world/canada/election-carney-liberal-party.html"><p><em>This article was updated after The Times learned that a remark attributed to Pierre Poilievre, the Conservative leader, was in fact an A.I.-generated summary of his views about Canadian politics that A.I. rendered as a quotation. The reporter should have checked the accuracy of what the A.I. tool returned. The article now accurately quotes from a speech delivered by Mr. Poilievre in April. [...] He did not refer to politicians who changed allegiances as turncoats in that speech.</em></p></blockquote> <p class="cite">&mdash; <a href="https://www.nytimes.com/2026/04/14/world/canada/election-carney-liberal-party.html">New York Times Editors’ Note</a></p> <p>Tags: <a href="https://simonwillison.net/tags/ai-ethics">ai-ethics</a>, <a href="https://simonwillison.net/tags/hallucinations">hallucinations</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/new-york-times">new-york-times</a>, <a href="https://simonwillison.net/tags/journalism">journalism</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a></p>

Echo Wang / Reuters: Sources: Cerebras plans to raise its IPO price range from $115-$125 per share to $150-$160 per share, potentially raising ~$4.8B at the top of the new range  —  Cerebras Systems is set to raise the size and price of its initial public offering as soon as Monday, as demand …

Peter and Rita Millard lived in countries across Europe after they retired. In 2023, they decided that Chiang Mai, Thailand, is where they want to grow old. Amanda Goh/Business Insider. After retiring from their jobs in the UK, Peter and Rita Millard spent years living in France and Malta. They outgrew both places and made a final move to a care facility in Chiang Mai, Thailand, in 2023. Peter says living in an aged care facility is an "insurance policy" for whoever gets left behind. This as-told-to essay is based on a conversation with Peter Millard, a 78-year-old retiree living in Chiang Mai, Thailand, with his wife, Rita. It has been edited for length and clarity. My wife and I spent more than 20 years moving between countries, trying to find the right place to grow old. We met while working in the health service in Birmingham, UK, before retiring on health grounds in the late '90s. Not ready to stop working, I retrained as a podiatrist and ran my own practice for five years, during which we built a small holiday home in southern France. We visited as often as possible. Over time, we realized how much we loved life in France. The Millards had never been to Asia before they decided to move to Thailand. Amanda Goh/Business Insider. France, then Malta When I finally retired again in 2003, we moved to Lodève, a small town about 31 miles from Montpellier, settling into the holiday home we had built. Our house was a 10-minute drive from the town, with terraces on three sides and lovely views of the surrounding hills. Even in a town of around 7,000 people, we built a close circle of friends, including our French neighbors, who had three children. We soon got involved with the school run, horse-riding lessons, and anything else when their mom was busy and transport was required. But as we got older, we started to think more practically about what we could manage in the future. The care facility they visited in Chiang Mai had lush greenery and a huge lake, and it felt like living in a park. Amanda Goh/Business Insider. You get to an age where there's no public transport, there's a lot of land to look after, and you start to wonder if you'll be able to cope in 10 years' time. We didn't want to return to the UK, so in 2012, we moved to Malta instead. Life there was more convenient, and we quickly settled into our three-bedroom penthouse located centrally on the island. After seven years, though, we felt we needed another change. We had outgrown Malta, as it is a very small island, and eventually decided to move back to the UK. This story is part of our Retirement: Chiang Mai series, which follows people choosing to retire in northern Thailand for reasons ranging from affordability to lifestyle and care. Read more: An American couple left Texas to retire in Thailand. They focus on getting 'one thing done a day.' He burned out, moved to Thailand with his family, and found early retirement didn't suit him If you've taken a similar path, or are planning to, email this reporter at agoh@businessinsider.com. A brief stint back in the UK Instead of returning to the Midlands where we previously lived, we moved to a retirement complex on the south coast. It was a big mistake. When we arrived, new management had just taken over, and the place became very inflexible, particularly around meal times and seating, and the food was poor. We lasted under three months before moving out and renting an apartment, where we were much happier. Although the town was comfortable, the UK no longer felt like home after so many years abroad. I ended up having a triple heart bypass just as the pandemic was hitting the country. We were trapped in England for nearly three years. By the end of a weeklong scouting trip, they made up their minds to move. In their minds, this would be where they'll grow old. Amanda Goh/Business Insider. Even so, the idea of living abroad again never quite left our minds. One Sunday, while reading the papers, Rita came across an article about a retirement resort in Chiang Mai, in northern Thailand. We had never been to that part of the world. But we both read the piece and said, "How about that?" We have no kids, so it's just the two of us. Chiang Mai, at last We arranged a visit to the Care Resort Chiang Mai, located on the outskirts of the city, to see if it could work for us. The couple said they were given free rein to decorate and make the space their own, helping it feel more like home. Amanda Goh/Business Insider. The surroundings were lovely, just like living in a park. We spent a few days at the resort and a few days in the city. By the end of the weeklong trip, we felt confident we had found the right place. Several months later, in February 2023, we moved to Chiang Mai. This will be our final move. We've done a lot of traveling since we've been together for 40-odd years, including on cruises. My wife says she never wants to see another airport. We were encouraged to make our villa our home, so we have refurnished and even changed the curtains. The couple said they want to live as independently as possible, while still having support available if needed. Amanda Goh/Business Insider. As we're fit and well, we want to carry on as normal, so we do most of the housework and prepare all our meals. We do make use of some of the facility's support when needed. Staff help with practical matters, such as visa renewals, and can arrange medical appointments for us. They see living in a care community as an "insurance policy" for whoever is left behind. Amanda Goh/Business Insider. Once you experience the kindness of the Thai people, you'll realize that this is a lovely country to spend your remaining years. In a way, choosing to live in a care community is about planning ahead. The idea was that whoever was left would be looked after by the people here. It's a bit like an insurance policy. We don't need any support now, but in the future, we may. Read the original article on Business Insider